The Cheapest Watt Is the One You Do Not Spend Twice

AI's energy problem is usually told as a story about scale: bigger models, more GPUs, new data centers. There is a quieter waste hiding underneath it. To trust an AI result, the standard answer is to run it again. Full re-execution doubles the energy. It does not have to.

The headline numbers about AI and electricity are real. Training and serving large models draws enormous power, data-center construction is straining grids, and the industry's answer so far has been to build more capacity. That conversation is important, but it skips a smaller, stranger inefficiency that sits inside the way we establish trust in a computation.

Here is the mechanism. When a regulated business needs assurance that an AI result is correct, the most common technical approach is redundancy: run the same job a second time, on a second machine, and check that the answers match. It is the honest instinct. It is also how you double the energy bill for that job. Every verified inference costs twice, because verification meant recomputation.

That is the waste worth attacking first, because it is pure overhead. The first run does the useful work. The second run produces nothing new except confidence. If you could get the confidence without the second full run, you would cut the energy cost of verified AI without touching a single model or buying a single GPU.

Proof does not have to mean redoing everything

You can establish that a large computation was performed correctly without repeating all of it. This is not exotic. Auditors do not re-add every number in a company's books; they sample, and the sampling is designed so that cheating is caught with high probability. The same statistical logic applies to computation. Check a well-chosen fraction of the work, make the choice unpredictable to the party being checked, and the math makes cutting corners a losing bet, at a fraction of the compute the full redo would cost.

The energy consequence is the whole point. Instead of paying 100 percent extra to verify, you pay a small fraction extra. The exact figure depends on the workload, but the direction is not in question: probabilistic verification turns the energy overhead of trust from a doubling into a rounding error. For a company running large batch inference jobs, that is the difference between two data-center bills and one.

The other half: the compute is already on and plugged in

There is a second lever, and it is even less intuitive. A large share of the world's most efficient compute is sitting idle in people's homes, already powered, already paid for. Modern phones are extraordinarily capable, and they draw a few watts, not a few hundred. A cloud inference GPU can pull 300 to 700 watts under load. A phone doing useful work draws on the order of single-digit watts.

This is not a thought experiment. A University of Zurich study benchmarked serverless compute delivered by a fleet of phones against the major cloud providers.

Put the two numbers side by side and the energy argument becomes concrete. The workload is the same. What changes is where it runs and what it draws.

The honest caveat matters here, and it is worth stating plainly rather than hiding. Phones are not the right executor for everything. Sustained, heavy generation on a phone runs into thermal limits: the device throttles, and a 2026 edge-inference study documented exactly this on recent flagship hardware. The place the phone fleet wins cleanly is large, asynchronous batch work that is not latency-sensitive: embedding a hundred thousand documents overnight, for example, where each device does a modest slice and the marginal energy cost is near zero. For that class of workload, which is a large and growing share of real production inference, the residential phone fleet is not a compromise. It is the more efficient machine.

Why the trust layer is what unlocks it

Here is the part that connects the two levers, and it is the reason this is an infrastructure argument and not just an efficiency tip. You cannot run regulated inference on a stranger's phone unless you can prove the result is correct. The whole reason serious workloads stay on expensive, centralized, high-draw hardware is that the buyer can point to a known operator and, in effect, buy the reputation instead of checking the work.

Verification is what removes that constraint. Once a result carries proof that a third party can check, without trusting the device it ran on, the physical location of the compute stops mattering for trust. The job can move to whatever hardware is cheapest and cleanest, because correctness is established by the proof, not by the machine. The trust layer is the thing that lets inference migrate off the 300-watt GPU and onto the 5-watt device that was already running.

The energy revolution is not building more. It is spending less, twice over.

The dominant plan for AI's energy demand is to expand: more data centers, more grid capacity, more construction. That plan treats demand as fixed and supply as the variable. There is another path that treats the waste as the variable.

Stop paying twice to verify, and the energy cost of trusted AI falls toward the cost of running it once. Let verified work migrate to idle, plugged-in, low-draw devices, and a large class of inference stops requiring new industrial capacity at all. Neither move needs a single new data center. Both need one thing: a way to prove a result is correct without re-running it and without trusting the hardware it ran on.

That is the quiet lever underneath the loud energy story. The cheapest watt in AI is not the one from a cleaner grid. It is the one you were about to spend a second time, and did not have to.

Try the live demo · Follow on LinkedIn · RSS